Actor-Partner Interdependence and Multilevel Approaches
Kenny, D. A., Kashy, D. A., & Cook, W. L. (2006). Dyadic data analysis. Guilford Press.
Visit: “David Kenny web page”
The dyad is the unit of interpersonal interaction
There is a variable that can be used to differentiate between the two members of the dyad.
Husband and wife
First and second author
The advisor and PhD. Student
There is not a variable that can be used to differentiate between the two persons.
Twins
Best friends (mutually chosen)
Roommates
In pairs, classify each example as Distinguishable (🧟️🧛️ = D) or Indistinguishable (🧟️🧟️ + I)
A therapist and their patient Two siblings Two co-authors who contributed equally A parent and an adolescent child Two romantic partners in a same-sex relationship
If the two scores from the two members of the dyad are non-independent, then those two scores are more similar to (or different from) from one another than are two scores from two people who are not members of the same dyad.
More similar within the dyad than the average
🧟🧟️
🧟️🧛️
🧛️🧛️
Activity: Spotting Non-Independence
Is there non-independence? If so, would scores be more similar or more different within dyads?
Married couples rating their satisfaction with the relationship.
Roommates reporting hours of sleep per night.
Siblings reporting their parents’ parenting style.
Warning
Non-independence is not always positive! In competitive dyads (e.g., debate opponents), members’ scores may be negatively correlated. The statistical problem of non-independence applies regardless of the direction.
Here is a simple dyadic dataset in long format:
Use pivot_wider() to reshape dyad_long into wide format where each row is one dyad and columns are satisfaction_A and satisfaction_B:
Here is a dyadic dataset in wide format (one row per dyad):
Use pivot_longer() to reshape dyad_wide into long format where each row is one partner:
When do you need each format?
lmer), most ggplot2 visualizationsQuick Poll
Imagine you run a regular regression (lm()) predicting relationship satisfaction from communication quality, using data from 50 couples (100 individuals). What could go wrong?
Sample: 10 cohabiting heterosexual couples (20 individuals)
Research question: Is household contribution associated with expectations about the relationship’s future?
Outcome: Likelihood of marriage within 5 years (%)
Predictor: Household contribution (composite of financial contribution + household labor, mean-centered)
Moderator: Cultural background
Why is this dyadic?
Each couple shares a household, finances, and a relationship. One partner’s contribution may affect both partners’ expectations about the future.
| Variable | Description | Coding |
|---|---|---|
Dyad |
Couple identifier | 1–10 |
Person |
Individual within the dyad | 1 = Person A, 2 = Person B |
Future |
Perceived likelihood of marriage within 5 years | 0–100% |
Gender |
Gender of the individual | -1 = Women, 1 = Men |
Contribution |
Household contribution (mean-centered) | Negative = below average |
Culture |
Cultural background of the couple | 1 = American, -1 = Asian |
Note on coding
All categorical predictors use effect coding (-1, 1) so that intercepts represent the grand mean and interactions are readily interpretable.
| Dyad | Person | Future | Contribution |
|---|---|---|---|
| 1 | 1 (W) | 75 | -10 |
| 1 | 2 (M) | 90 | -5 |
| 2 | 1 (W) | 55 | 0 |
| 2 | 2 (M) | 75 | 10 |
| 3 | 1 (W) | 45 | -10 |
| 3 | 2 (M) | 33 | -15 |
We can progressively build complexity:
Why is this wrong?
Standard lm() treats all 20 observations as independent, ignoring the dyadic structure.
We add a random intercept for Dyad to account for non-independence within couples:
Is the random effect (the dyadic clustering) significant?
What to look for
A significant p-value means the dyadic clustering matters — the random intercept improves model fit beyond the OLS model.
How many random effects are in this model?
Two random factors:
First, fit an empty model (no predictors) to isolate the dyadic variance:
Now compute the ICC manually:
(1 | Dyad)Try it yourself
Go back to the plot code and change facet_wrap(~Culture_label) to facet_wrap(~factor(Dyad)) to see each couple separately.
Actor Partner Interdependece Model
Answer the following for a dyadic relationship of your choice:
Who is the dyad? (e.g., parent-child, romantic partners, coach-athlete) Distinguishable or indistinguishable? Why? What is X (the predictor)? (e.g., stress, communication style) What is Y (the outcome)? (e.g., well-being, performance) What would a significant actor effect mean in plain language? What would a significant partner effect mean? Which effect do you hypothesize to be larger? Why?
Confirmatory Factor Analysis
Both members of the dyad answer the same items. Because they share a context (the same family), the residuals of matching items are expected to be correlated.
A dyadic CFA accounts for this by:
Our example
Dyad: Parent and adolescent child
Construct: Perceived harm of alcohol use in young people
Items (same 4 for both):
We simulate 200 parent-child dyads answering 4 items each:
Mplus expects a space-delimited .dat file with no headers:
To save the file on your computer, run this in R (not webR):
File: dyadic_cfa.inp
Download this file from the course website and run it in Mplus.
TITLE: Dyadic CFA - Perceived Harm of Alcohol
Parent-Child Dyads (200 families)
DATA: FILE = "dyadic_cfa.dat";
VARIABLE:
NAMES = C1 C2 C3 C4 P1 P2 P3 P4;
USEVARIABLES = C1 C2 C3 C4 P1 P2 P3 P4;
MODEL:
! Latent factors
CHILD BY C1* C2 C3 C4;
PARENT BY P1* P2 P3 P4;
! Fix factor variances for identification
CHILD@1;
PARENT@1;
! Factor correlation (dyadic association)
CHILD WITH PARENT;
! Correlated residuals across matching items
C1 WITH P1;
C2 WITH P2;
C3 WITH P3;
C4 WITH P4;
OUTPUT: STDYX MODINDICES CINTERVAL;C1 WITH P1; lines and re-run. How does model fit change?To test whether the items function the same way for parents and children, constrain the loadings to be equal:
Compare models
Run the free and constrained models. A non-significant chi-square difference means the items work equivalently for both parents and children — supporting measurement invariance.

EPH 752 Advanced Research Methods